latentbrief
Back to news
Launch3w ago

Amazon Bedrock Introduces Programmatic Tool Calling and Custom Evaluators

AWS ML Blog1 min brief

In brief

  • Amazon Bedrock has unveiled a new approach to how large language models (LLMs) interact with external tools.
  • Programmatic tool calling (PTC) allows models to write code, typically in Python, that can invoke multiple tools programmatically within a sandboxed environment.
    • This method reduces latency and token usage by executing all intermediate results outside the model's context window.
  • For instance, tasks like analyzing thousands of expense records become more efficient as only the final result returns to the model.
  • Additionally, Bedrock now offers custom code-based evaluators through AWS Lambda functions.
    • These evaluators enable precise checks for structured data, such as JSON outputs from APIs or financial metrics.
  • By using deterministic code, users can validate tool responses without relying on costly LLM evaluations.
    • This feature is particularly useful in domains like finance, where accuracy and compliance are critical.
    • These advancements aim to streamline development workflows and improve the reliability of agentic applications.
  • As more tools become available, developers can expect even greater efficiency and control over their AI-driven systems.

Terms in this brief

Programmatic Tool Calling
A method where large language models write code to invoke tools programmatically, improving efficiency by executing intermediate results outside the model's context window. This reduces latency and token usage, making tasks like analyzing expense records more efficient.

Read full story at AWS ML Blog

More briefs